Fuzzy Sets and Systems 126 (2002) 389–399

www.elsevier.com/locate/fss

Fuzzy least-squares linear regression analysis for fuzzy

input–output data
Miin-Shen Yang ∗ , Tzu-Shun Lin
Department of Mathematics, Chung-Yuan Christian University, Chung-Li, 32023 Taiwan, ROC

Received 27 July 1997; received in revised form 7 September 2000; accepted 12 February 2001

Abstract
A fuzzy regression model is used in evaluating the functional relationship between the dependent and independent
variables in a fuzzy environment. Most fuzzy regression models are considered to be fuzzy outputs and parameters but
non-fuzzy (crisp) inputs. In general, there are two approaches in the analysis of fuzzy regression models: linear-programming-
based methods and fuzzy least-squares methods. In 1992, Sakawa and Yano considered fuzzy linear regression models
with fuzzy outputs, fuzzy parameters and also fuzzy inputs. They formulated multiobjective programming methods for the
model estimation along with a linear-programming-based approach. In this paper, two estimation methods along with a
fuzzy least-squares approach are proposed. These proposed methods can be e ectively used for the parameter estimation.
Comparisons are also made between them.
c 2002 Elsevier Science B.V. All rights reserved.

Keywords: Fuzzy linear regression; Estimation; Multiobjective programming; Fuzzy least-squares; Clusterwise regression;
Noise cluster; Robustness

1. Introduction

Fuzzy regression analysis is a fuzzy (or possibility) type of classical regression analysis. It is used in eval-
uating the functional relationship between the dependent and independent variables in a fuzzy environment.
Tanaka et al. [11] initiated a study in fuzzy linear regression analysis which considered the parameter estima-
tion of models as linear programming problems. Subsequently, there have been many other studies which can
be roughly divided into two approaches: linear-programming-based methods [8– 11] and fuzzy least-squares
methods [1,5,14]. Most of these fuzzy regression models are considered with fuzzy outputs and fuzzy pa-
rameters but non-fuzzy (crisp) inputs. In this paper fuzzy linear regression models with fuzzy outputs, fuzzy
parameters, and fuzzy inputs are considered.

∗Corresponding author. Tel.: +886-3-456-3171; fax: +886-3-456-3160.

E-mail address: [email protected] (M.-S. Yang).

0165-0114/02/$ - see front matter  c 2002 Elsevier Science B.V. All rights reserved.
PII: S 0 1 6 5 - 0 1 1 4 ( 0 1 ) 0 0 0 6 6 - 5
390 M.-S. Yang, T.-S. Lin / Fuzzy Sets and Systems 126 (2002) 389–399

Sakawa and Yano [8] proposed a fuzzy parameter estimation for the fuzzy linear regression (FLR) model
Yj = A0 + A1 Xj1 + · · · + Ak Xjk ; j = 1; : : : ; n;
where both input data Xj1 ; : : : ; Xjk and output data Yj are fuzzy, by using three indices for the equalities
between two fuzzy numbers M and N as

Pos(M = N ) = sup min{M (x); N (x)};

x

Nes(M ⊂ N ) = inf max{1 − M (x); N (x)};

x

Nes(M ⊃ N ) = inf max{M (x); 1 − N (x)};

x

where M (x) and N (x) are membership functions of M and N , and Pos and Nes are short for Possibility
and Necessity. Then three types of multiobjective programming problems were formulated for the parameter
estimation of FLR models along with a linear-programming-based approach. This multicriterial analysis of FLR
models provided a good method of parameter estimation by using the vagueness of the model via some indices
of inclusion relations. Alternatively, a fuzzy least-squares approach directly uses information included in the
input–output data set and considers the measure of best tting based on distance under fuzzy consideration.
Fuzzy least-squares are fuzzy extensions of ordinary least-squares. In this paper two types of fuzzy least-squares
are proposed as the parameter estimation for the FLR model Yj = A0 + A1 Xj1 + · · · + Ak Xjk ; j = 1; : : : ; n. These
proposed methods have proved to be a good alternative approach to a multiobjective programming approach.
Section 2 presents the rst method, called approximate-distance fuzzy least-squares. In order to treat hetero-
geneous data sets and detect outliers, a clusterwise fuzzy least-square with a noise cluster is also presented.
In Section 3, another fuzzy least-squares method, called an interval-distance fuzzy least-square, is proposed.
Numerical results are presented in Section 4. Finally, conclusions are made in Section 5.

2. Approximate-distance fuzzy least-squares

The following FLR model is considered:

Yj = A0 + A1 Xj1 + · · · + Ak Xjk ; j = 1; : : : ; n; (2.1)
where outputs Yj = (myj ; yj ; yj )LR , inputs Xji = (mxji ; xji ; xji )LR and parameters Ai = (mai ; ai ; ai )LR ; i =
1; : : : ; k; j = 1; : : : ; n so that the notion M = (m; ; )LR is an LR-type fuzzy number with its membership
function
  
m−x
L for x 6 m




M (x) =   ( ¿0; ¿0):
 x−m
R for x ¿ m



In M = (m; ; )LR ; m is called the mean value, and and are called the left and right spreads, respectively.
If L(x) = R(x) = 1 − x, then M = (m; ; )LR is called a triangular fuzzy number, denoted by M = (m; ; )T .
The diculty in treating model (2.1) of fuzzy input–output data is that Ai Xji may not be of LR-type.
Although the product of two LR-type fuzzy numbers may not be an LR-type fuzzy number, Dubois and Prade
[6] presented an approximation form. Therefore, Yang and Ko [14] suggested an approximation type of fuzzy
least-squares for an extension to their results. In this section, the idea of approximation is used to present an
algorithm for parameter estimation of the FLR model (2.1).
 M.-S. Yang, T.-S. Lin / Fuzzy Sets and Systems 126 (2002) 389–399 391

Dubois and Prade [6] presented the approximation formula as follows:

For M = (m; ; )LR and N = (n; ; )LR ,
(1) if M ¿0 and N ¿0, then

(m; ; )LR (n; ; )LR ≃ (mn; m + n ; m + n )LR ;

(2) if M ¡0 and N ¿0, then

(m; ; )LR (n; ; )LR ≃ (mn; n − m; n − m )LR ;

(3) if M ¡0 and N ¡0, then

(m; ; )LR (n; ; )LR ≃ (mn; −n − m; −n − m )RL :

Assume that

H1 = {Ap | Ap ¿ 0; p ∈ {1; : : : ; k}};

H2 = {Ap | Ap ¡ 0; p ∈ {1; : : : ; k}};

sjp = 1 if Xjp ¿ 0 and 0 if Xjp ¡ 0:

Then A0 + A1 Xj1 + · · · + Ak Xjk ≃ (m̃j ; ˜ j ; ˜j )LR , where

k

m̃j = ma0 + (map mxjp )
p=1

˜j = a0 + [sjp (map xjp + mxjp ap ) + (1 − sjp )(map xjp − mxjp ap )]
Ap ∈H1


+ [sjp (mxjp ap − map xjp ) + (1 − sjp )(−map xjp − mxjp ap )];
Ap ∈H2


˜ = a0 + [sjp (map xjp + mxjp ap ) + (1 − sjp )(map xjp − mxjp ap )]
j
Ap ∈H1


+ [sjp (mxjp ap − map xjp ) + (1 − sjp )(−map xjp − mxjp ap )]:
Ap ∈H2

2
Since (A0 + A1 Xj1 + · · · + Ak Xjk ) is of approximate LR-type, the distance dLR de ned on two LR-type fuzzy
numbers by Yang and Ko [14] is used here so that

d2LR (X; Y ) = (mx − my )2 + [(mx − l x ) − (my − l y )]2 + [(mx + r x ) − (my + r y )]

2
;
1 1
where X = (mx ; x; x )LR ; Y = (my ; y ; y )LR and l = 0
L−1 (!) d!, r = 0
R−1 (!) d!.
392 M.-S. Yang, T.-S. Lin / Fuzzy Sets and Systems 126 (2002) 389–399

Thus, the following objective function is considered:

n

J (A0 ; A1 ; : : : ; Ak ) = d2LR (Yj ; A0 + A1 Xj1 + · · · + Ak Xjk )
j=1

n

= [(myj − m̃j )2 + ((myj − l yj ) − (m̃j − l ˜j ))2 + ((myj + r yj ) − (m̃j + r ˜j ))2 ]:
j=1

The minimization of J (A0 ; A1 ; : : : ; Ak ) over Ai subject to ai ¿ 0 and ai ¿ 0; i = 0; 1; : : : ; k is called an appro-

ximate-distance fuzzy least-squares method.
The regression analysis in the case of heterogeneity of observations is commonly presented in actual
practice. In order to overcome the heterogeneous problem, an idea of clusterwise fuzzy regression anal-
ysis [14] is adopted here. Fuzzy clustering, which has been widely studied and applied in a variety of
substantive areas [2,12,13,15], is used to treat the heterogeneous problem. Clusterwise fuzzy regression
embeds fuzzy clustering into fuzzy regression model tting at each step in the iterations. Given a data set
{(Xj1 ; : : : ; Xjk ; Yj ); j = 1; : : : ; n}, a clusterwise FLR model is tted to the data set
Yj = A0i + A1i Xj1 + · · · + Aki Xjk ; i = 1; : : : ; c; j = 1; : : : ; n:
c
Let ij ∈ [0; 1] with i=1 ij = 1 for all j = 1; : : : ; n. The notation ij is used to represent the membership of
the jth data point (Xj1 ; : : : ; Xjk ; Yj ) belonging to the ith class. After embedding ij to the objective function J ,
one has a clusterwise objective function
c 
 n
J (; A0i ; : : : ; Aki ) = ijm d2LR (Yj ; A0i + A1i Xj1 + · · · + Aki Xjk );
i=1 j=1

where m ¿ 1 is the index of fuzziness. According to the Lagrange multiplier, one has the necessary condition
for ij with
 −1
c 2 1=(m−1) i = 1; : : : ; c;
 (dLR (Yj ; A0i + A1i Xj1 + · · · + Aki Xjk ))
ij =  2 1=(m−1)
 ;
p=1
(dLR (Yj ; A0p + A1p Xj1 + · · · + Akp Xjk )) j = 1; : : : ; n:

Outliers always have immense e ects in model tting, especially in regression. E.g., they decrease the
accuracy of estimation. Thus, robust regression and outlier detection become an important consideration (see
[7]). Robustness seems to be more important in fuzzy regression. However, our clusterwise fuzzy regression
gives the restriction of membership functions with i ij = 1 for j = 1; : : : ; n so that the results will be
deteriorated due to outliers and noise. An easy way to modify this clusterwise fuzzy regression to detect and
tolerate noise and outliers is to use the idea of Dave’s noise cluster [3]. A noise cluster is a cluster which
contains noise points or outliers so that all the points have equal a priori probability of belonging to a noise
cluster. Assume that the cluster (c + 1) is a noise cluster. Then the objective function becomes
c+1 
 n
0
J (; A0i ; : : : ; Aki ) = ijm d2ij ; (2.2)
i=1 j=1

where
d2LR (Yj ; A0i + A1i Xj1 + · · · + Aki Xjk ); i = 1; : : : ; c; j = 1; : : : ; n;

d2ij =
2 ; i = c + 1; j = 1; : : : ; n;
 M.-S. Yang, T.-S. Lin / Fuzzy Sets and Systems 126 (2002) 389–399 393

c n
 
i=1 j=1 d2ij
2 =  ;  ¿ 0 a constant:
nc

The necessary condition of minimization of J 0 over  is

 −1
c+1
 (d2ij )1=(m−1)
ij =  2 )1=(m−1)
 ; i = 1; : : : ; c + 1; j = 1; : : : ; n: (2.3)
p=1
(dpj

Note that a connection between the noise cluster approach and robust statistics had been discussed by Dave
and Krishnapuram [4]. Thus, a robust clusterwise fuzzy regression algorithm for fuzzy input–output data is
obtained as follows:

Algorithm 1.
S(1) Given m¿1 and c ∈ {1; 2; : : : ; n − 1}.
S(2) Choose initials ij(0) ; i = 1; : : : ; c + 1; j = 1; : : : ; n.
S(3) Find the approximate minimizers Â0i ; Â1i ; : : : ; Âki of J 0 in (2.2).
S(4) Calculate ij(0) by using (2.3).
S(5) Compare ij(0) to ij(1) , using any convenient norm. If ij(1) is suciently close to ij(0) , stop. Otherwise,
set ij(0) = ij(1) and go to (S3).
Here the approximate minimizers Aˆ0i ; Aˆ1i ; : : : ; Aˆki of J 0 can be obtained by using computer software such as
S-plus, Mathematica, etc. Algorithm 1 is called an approximate-distance fuzzy least-squares algorithm.

3. Interval-distance fuzzy least-squares

The fuzzy least-squares method presented in Section 2 uses an approximation formula for the product of
two LR-type fuzzy numbers. It obviously has the drawback that if the di erence between L and R or the
spreads and is large, then the approximation becomes inferior. In this section another fuzzy least-squares
method is presented.
Since A0 + A1 Xj1 + · · · + Ak Xjk may not be of LR-type, Sakawa and Yano [8] presented its !-level set with
a closed interval so that

(A0 + A1 Xj1 + · · · + Ak Xjk )! = (Ã ⊗ X̃ j )! = [(Ã ⊗ X̃ j )L! ; (Ã ⊗ X̃ j )R! ];

where
k

(Ã ⊗ X̃ j )L! = {min{(mai − L−1
ai (!) ai )(mxji − L−1
xji (!) xji ); (mai − L−1
ai (!) ai )(mxji + R−1
xji (!) xji )}};
i=0

k

(Ã ⊗ X̃ j )R! = {max{(mai + R−1
ai (!) ai )(mxji + R−1
xji (!) xji ); (mai + R−1
ai (!) ai )(mxji − L−1
xji (!) xji )}}:
i=0

Here (M )! denotes the !-level set of the fuzzy number M . It is known that the fuzzy output Yj has its
!-level set (Yj )! = [YjwL ; YjwR ] with
L R
Yjw = myj − L−1
yj (!) yj ; Yjw = myj + R−1
yj (!) yj :
394 M.-S. Yang, T.-S. Lin / Fuzzy Sets and Systems 126 (2002) 389–399

In order to create a fuzzy least-squares objective function, a distance D 2 (assume that D 2 exists) between Yj
and A0 + A1 Xj1 + · · · + Ak Xjk is de ned as
 1
D2 (Yj ; A0 + A1 Xj1 + · · · + Ak Xjk ) = L
((Yj! − (Ã ⊗ X̃ j )L! )2 + (Yj!
R
− (Ã ⊗ X̃ j )L! )2 ) d!:
0

Thus, the interval-distance fuzzy least-squares objective function  is de ned as

n

(A0 ; A1 ; : : : ; Ak ) = D2 (Yj ; A0 + A1 Xj1 + · · · + Ak Xjk )
j=1

n 
 1
= L
((Yj! − (Ã ⊗ X̃ j )L! )2 + (Yj!
R
− (Ã ⊗ X̃ j )L! )2 ) d!:
j=1 0

In order to treat min and max operators in (A˜ ⊗ X̃j )! , the parameters A0 ; A1 ; : : : ; Ak are divided into the
following three groups:

P1 = {Ai | mai − L−1

ai (!) ai ¿ 0};

P2 = {Ai | − R−1
ai (!) ai 6 mai ¡ L−1
ai (!) ai };

P3 = {Ai | mai ¡ −R−1

ai (!) ai }:

Thus, (A˜ ⊗ X̃j )L! and (A˜ ⊗ X̃j )R! in the objective function (A0 ; A1 ; : : : ; Ak ) can be rewritten as
 
(Ã ⊗ X̃ j )L! = (mai − L−1
ai (!) ai )(mxji − L−1
xji (!) xji ) + (mai − L−1
ai (!) ai )(mxji + R−1
xji (!) xji );
Ai ∈P1 Ai ∈P2 ∪P3

(Ã ⊗ X̃ j )R! = (mai + R−1
ai (!) ai )(mxji + R−1
xji (!) xji );
Ai ∈P1 ∪P2

+ (mai + R−1
ai (!) ai )(mxji − L−1
xji (!) xji ):
Ai ∈P3

Similarly, the objective function (A0 ; A1 ; : : : ; Ak ) is extended to a robust-type objective function

c+1 
 n
0 (; A0i ; : : : ; Aki ) = ijm d2ij ; (3.1)
i=1 j=1

where

D2 (Yj ; A0i + A1i Xj1 + · · · + Aki Xjk ); i = 1; : : : ; c; j = 1; : : : ; n;


d2ij =
2 ; i = c + 1; j = 1; : : : ; n;
c n
 
2 i=1 j=1 d2ij
 = ;  ¿ 0 a constant:
nc
 M.-S. Yang, T.-S. Lin / Fuzzy Sets and Systems 126 (2002) 389–399 395

The necessary condition of minimization of 0 over  is

 −1
c+1
 (d2ij )1=(m−1)
ij =   ; i = 1; : : : ; c + 1; j = 1; : : : ; n: (3.2)
(d2 )1=(m−1)
p=1 pj

Similar to Algorithm 1, an algorithm through (3.1) and (3.2) is created and called an interval-distance fuzzy
least-squares algorithm.
We are now interested in comparisons of interval-distance and approximate-distance fuzzy least-squares.
Both methods are for optimization based on fuzzy least-squares objective functions. They are also sensitive
to outliers and noise, but can be improved by using the noise cluster approach to have a high ability to
detect noise. Of course, they are di erent in that one uses an approximation formula for fuzzy product of two
LR-type fuzzy numbers and the other uses the integration of interval distance to their !-level sets. It is known
that if the di erence between L and R or the spreads and is large, then the approximate-distance method
obviously becomes inferior to the interval-distance method. However, both methods can get very close results
when the spreads and are small.
Here let us discuss the case of L = R. Assume that B = (mb ; b ; b )LL ¿0 and X = (mx ; x ; x )LL ¿0. Then
the approximation formula has

BX ≃ (mb mx ; mb x + mx b ; mb x + mx b )LL

so that its !-level set is

(BX )A! = [mb mx − L−1 (!)(mb x + mx b ); mb mx + L−1 (!)(mx x + mx b )]:

The interval presentation gives its !-level set

(BX )I! = [(mb − L−1 (!) b )(mx − L−1 (!) x ); (mb + L−1 (!) b )(mx + L−1 (!) x )]:

In this case, the interval (BX )I! is always located to the right of (BX )A! for all ! with left-side di er-
ence = (L−1 (!))2 b x , and right-side di erence = (L−1 (!))2 b x .
If the left spreads b and x (or b and x ) are small then left (or right) end points of both intervals are
close together. Similarly, we can analyze the cases of (B¡0; X ¿0) and (B¡0; X ¡0).

4. Numerical results

In Sections 2 and 3 two fuzzy least-squares methods have been constructed for the estimation of an FLR
model with fuzzy input–output data. In this section some numerical examples are given.

Example 1. In this example, the fuzzy input–output data from Sakawa and Yano [8] are used. The data set
is shown in Table 1, and the results are shown in Table 2.

From the results in Table 2, it is seen that the parameter estimates and sum of squares of residuals (SSR)
from the approximate-distance and interval-distance fuzzy least-squares methods are almost the same. This is
obviously because L = R and the spreads are small. According to Sakawa and Yano [8], the threshold  is
subjectively speci ed by the decision maker. If the decision maker is satis ed with the current values of the
fuzziness and , then stop. Otherwise, the decision maker must update the fuzzy threshold  by considering
the current values of the fuzziness and  together with the adjustment rate between the fuzziness and , and
then calculate again.
396 M.-S. Yang, T.-S. Lin / Fuzzy Sets and Systems 126 (2002) 389–399

Table 1
Fuzzy input–output data set [8] X = (mx ; x ; x )T ; Y = (my ; y ; y )T

No. mx x (= x) my y (= y)

1 2.0 0.5 4.0 0.5

2 3.5 0.5 5.5 0.5
3 5.5 1.0 7.5 1.0
4 7.0 0.5 6.5 0.5
5 8.5 0.5 8.5 0.5
6 10.5 1.0 8.0 1.0
7 11.0 0.5 10.5 0.5
8 12.5 0.5 9.5 0.5

Table 2
Parameter estimates and SSR for di erent methods

Approximate Interval Sakawa and Yano (CP1 )

distance distance
 = 0:1 0.2 0.3 0.4 0.5 0.6 0.7

m̂a 3.4967 3.5840 3.3770 3.3000 3.2220 3.1454 3.2010 3.3670 3.5080
ˆa (= ˆa ) 0.2915 0.2916 0 0 0 0 0.1703 0.4260 0.9242
m̂b 0.5293 0.5190 0.5641 0.5709 0.5779 0.5848 0.5788 0.5595 0.5429
ˆb (= ˆb ) 0.0042 0.0005 0.0207 0.0318 0.0463 0.0657 0.0806 0.1108 0.1530
SSR based on
approximate-distance
2 (Y; Yˆ )
dLR 15.6020 15.7760 16.3535 16.4744 16.7900 17.4150 18.3913 21.3970 31.3416
SSR based on
interval-distance
D2 (Y; Yˆ ) 11.0340 10.4754 11.1688 11.1606 11.3375 11.8460 12.9970 16.8760 29.4790

It is unnecessary to consider the threshold  for the proposed fuzzy least-squares methods. This is because
 has been taken into account in the de ned distance. Thus, the parameter estimates of fuzzy least-squares
methods are more simple and ecient than those of Sakawa and Yano [8] since the structure and information
encountered in the data set is taken into account at each iteration of fuzzy least-squares algorithms. On
the other hand, the results in Table 2 show that the SSRs based on the distance D 2 and dLR 2
do not change
signi cantly when the values of  in CP1 are between 0.1 and 0.5, but SSRs undergo an immense change when
the values of  are between 0.5 and 0.7. This indicates that the best value of  in algorithm CP1 can be selected
between 0.55 and 0.65 according to the SSRs based on the proposed distance, which matches the selected
value 0.6 of  according to the method of adjusting rates in Sakawa and Yano [8]. Therefore, the proposed
fuzzy least-squares methods can be also thought of as the -selection of Sakawa and Yano [8]. The data set
in Table 3 and the results in Table 4 demonstrate this more clearly. Obviously,  = 0:5 is selected for CP1
according to the SSRs based on the de ned distance.

Example 2. In this example, how the fuzzy least-squares and Sakawa and Yano methods are sensitive to
outliers and how the noise cluster is used to solve the drawback are demonstrated. Thus, a robust type of
fuzzy least-squares methods is presented.
 M.-S. Yang, T.-S. Lin / Fuzzy Sets and Systems 126 (2002) 389–399 397

Table 3
Fuzzy input–output data set X = (mx ; x ; x )T ; Y = (my ; y ; y )T

No. mx x (= x) my y (= y)

1 1 0.5 0.875 0.5

2 2 0.5 1.575 0.5
3 2.5 0.5 1.750 0.5
4 3 0.8 1.575 0.8
5 4 1.0 2.100 1.0
6 5 0.6 2.275 0.6
7 6.5 0.4 2.975 0.4
8 7.5 1.0 3.500 1.0
9 9 0.5 3.150 0.5
10 10 0.5 4.375 0.5
11 10.5 0.5 3.675 0.5
12 11 0.5 3.500 0.5

Table 4
Parameter estimates and SSR for di erent methods

Approximate Interval Sakawa and Yano (CP1 )

distance distance
 = 0:1 0.2 0.3 0.4 0.5 0.6 0.7

m̂a 0.8711 0.9096 1.1806 1.1175 1.0543 0.9912 0.9281 0.865 0.8018
ˆa ( = ˆa ) 0.3310 0.4565 0 0 0 0 0 0 0
m̂b 0.2898 0.2834 0.2625 0.2886 0.2747 0.2809 0.2871 0.2935 0.2998
ˆb (= ˆb ) 0.0015 0 0.0001 0.0077 0.0175 0.0307 0.0493 0.0775 0.1254
SSR based on
approximate-distance
2 (Y; Yˆ )
dLR 3.9702 4.0420 5.2651 4.8048 4.4318 4.1770 4.1304 4.5509 6.3760
SSR based on
interval-distance
D2 (Y; Yˆ ) 2.7398 2.6660 4.9122 4.3678 3.8617 3.4250 3.1550 3.3600 5.1880

The data set in Table 1 is used, but the fth point my = 8:5 is changed to my = 18:5. It is seen that the fth
data point becomes an outlier. Then the algorithms are implemented on this data set, with the results shown
in Table 5. It is found that the SSRs using algorithms without a noise cluster increase signi cantly to more
than 200; but if fuzzy least-squares algorithms are used with a noise cluster, then the SSRs decrease to less
than 88. On the other hand, it is found that the membership of the fth data point belonging to the noise
cluster is 0.92. Clearly, the point is detected to be an outlier. The proposed fuzzy least-squares algorithms
with a noise cluster are a robust type and also an outlier detector.

5. Conclusions

Although there are many articles on fuzzy linear regression analysis, most of them discuss fuzzy linear
regression models with fuzzy outputs and fuzzy parameters but non-fuzzy (crisp) inputs. Sakawa and Yano
[8], who considered fuzzy linear regression models with fuzzy outputs, fuzzy parameters and also fuzzy inputs,
proposed multiobjective programming methods. In their algorithm, the decision maker needs to subjectively
398 M.-S. Yang, T.-S. Lin / Fuzzy Sets and Systems 126 (2002) 389–399

Table 5
Parameter estimates and SSR when the data my = 8:5 in Table 1 changes to my = 18:5

Approximate-distance Interval-distance Sakawa and Yano (CP1 )

 = 0:4  = 0:5  = 0:6

With a noise cluster

(yes or no) No Yes No Yes
m̂a 4.009 3.457 4.097 3.534 3.145 3.115 3.196
ˆa (= ˆa ) 0.230 0.202 0.320 0.261 0 0 0
m̂b 0.626 0.515 0.615 0.507 1.128 1.134 0.126
ˆb (= ˆb ) 0.004 0.005 0 0.003 0.971 1.191 1.526
SSR based on
approximate-distance
2 (Y; Yˆ )
dLR 306.74 83.68 305.78 88.03 856.15 994.95 1498.49
SSR based on
interval-distance
D2 (Y; Yˆ ) 204.08 55.47 204.04 54.41 770.01 956.16 1344.34

specify a threshold . It is known that a fuzzy least-squares approach is originally a fuzzy extension of
ordinary least-squares. This approach uses information included within the input–output data. Sakawa and
Yano’s methods, instead of the measure of best- tting, use indices for equalities or inclusion relationships
between fuzzy numbers. A fuzzy least-squares approach, not considering equalities or inclusion relationships,
directly uses the measure of best- tting under fuzzy consideration. In this paper alternative fuzzy least-squares
methods for fuzzy linear regression models with fuzzy input–output data have been proposed. It is unnecessary
to consider the threshold  for the proposed methods, since  has been taken into account at each iteration of
the algorithm according to the structure and information encountered in the data set. On the other hand, the
idea of a noise cluster has been applied and a robust type of fuzzy least-squares algorithms has been presented.
In conclusion, the proposed methods suggest an ecient alternative procedure to estimate parameters for fuzzy
linear regression models with fuzzy outputs, fuzzy parameters and also fuzzy inputs.

Acknowledgements

The authors would like to thank the anonymous reviewers for their constructive comments to improve the
presentation of the paper. This work was supported in part by the National Science Council of the Republic
of China under Grant NSC 87-2118-M-033-001.

References

[1] M. Albrecht, Approximation of functional relationships to fuzzy observations, Fuzzy Sets and Systems 49 (1992) 301–305.
[2] J.C. Bezdek, Pattern Recognition with Fuzzy Objective Function Algorithms, Plenum Press, New York, 1981.
[3] R.N. Dave, Characterization and detection of noise in clustering, Pattern Recognition Lett. 12 (1991) 657–664.
[4] R.N. Dave, R. Krishnapuram, Robust clustering method: a uni ed view, IEEE Trans. Fuzzy Systems 5 (1997) 270–293.
[5] P. Diamond, Fuzzy least squares, Inform. Sci. 46 (1988) 141–157.
[6] D. Dubois, H. Prade, Fuzzy Sets and Systems: Theory and Applications, Academic Press, New York, 1980.
[7] P.J. Rousseeuw, A.M. Leroy, Robust Regression and Outlier Detector, Wiley, New York, 1987.
[8] M. Sakawa, H. Yano, Multiobjective fuzzy linear regression analysis for fuzzy input–output data, Fuzzy Sets and Systems 47 (1992)
173–181.
 M.-S. Yang, T.-S. Lin / Fuzzy Sets and Systems 126 (2002) 389–399 399

[9] H. Tanaka, H. Ishibuchi, Identi cation of possibilistic linear systems by quadratic membership functions of fuzzy parameters, Fuzzy
Sets and Systems 41 (1991) 145–160.
[10] H. Tanaka, H. Ishibuchi, S. Yoshikawa, Exponential possibility regression analysis, Fuzzy Sets and Systems 69 (1995) 305–318.
[11] H. Tanaka, S. Vegima, K. Asai, Linear regression analysis with fuzzy model, IEEE Trans. Systems Man Cybernet. 12 (1982)
903–907.
[12] M.S. Yang, A survey of fuzzy clustering, Math. Comput. Modelling 18 (1993) 1–16.
[13] M.S. Yang, On a class of fuzzy classi cation maximum likelihood procedures, Fuzzy Sets and Systems 57 (1993) 365–375.
[14] M.S. Yang, C.H. Ko, On cluster-wise fuzzy regression analysis, IEEE Trans. Systems Man Cybernet. B 27 (1) (1997) 1–13.
[15] M.S. Yang, J.A. Pan, On fuzzy clustering of directional data, Fuzzy Sets and Systems 91 (1997) 319–326.